agentleFS
Sign inStart free
p10ns11yp10ns11y/skills/adversarial-audit/SKILL.md

adversarial-audit

Skill2 starsChanged 40 days agoMITView on GitHub

Adversarial verification of claimed work: read shipped modules (not memory), grep wiring and approvals, run tests, demand evidence before pass verdict. Use when reviewing agent output, PRs, or security-sensitive changes before merge. Triggers: adversarial audit, refute claims, evidence before verdict, pre-merge review.

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

Create a free account to mount
Files
SKILL.md

adversarial-audit

Load rule: Formal SoT. Distinct from re-running verify cmds alone (agent-orchestrator VERIFY) and from Outer REVIEW_GATE shape (control-graph) — this skill is refute-first evidence pack with independent reads.
CLT: ../rules/clt-dual-load.mdc — review is protected germane load; never collapse to implementer self-summary.

// Role
Reviewer ⊥ ImplementerContext   // fresh session/context; try to refute claims

// Axioms
A1  Default real=false until independent evidence
A2  Read shipped files + run commands — never trust summary alone
A3  Verdict requires: scope check + tests/cmds + explicit gaps
A4  Compose: Orch owns re-run of verify cmds; CG owns when REVIEW_GATE fires; this skill owns attack surface + evidence artifact
A5  Fail closed on missing evidence for security/auth/deps claims
A6  Evaluate(δ) ≔ (C, E, η)

Related: agent-orchestrator · control-graph · fix-dependency-security.

Use / skip

UseSkip
worker/agent claimed "done"pure design chat, no diff
pre-merge high stakestypo you wrote yourself 30s ago (still run lint if easy)
security / auth / deps changes
refute marketing claims in PR body

vs Orch VERIFY: Orch re-runs listed cmds and checks scope. This skill also opens files, greps wiring, hunts missing authz/tests, and writes a verdict artifact.
vs CG REVIEW_GATE: CG is phase/budget control. This skill is the review content procedure for the review Role.

Procedure

1 Scope inventory → 2 Read shipped code → 3 Run verify cmds → 4 Attack claims → 5 Verdict artifact

1. Scope inventory

git diff <base>...HEAD --stat
git log <base>..HEAD --oneline

List claimed outcomes from brief/PR body.

2. Read shipped code

Open every file that implements a claim; grep wiring:

rg -n "pattern|approval|TODO|FIXME|XXX" --glob '!**/node_modules/**'

3. Run verify commands

From brief / AGENTS.md / CI — capture exit codes:

# examples — use project actuals
pnpm type-check && pnpm lint
cargo test -q

4. Attack claims

  • Does diff match the claim?
  • Missing error paths, authz, cleanup?
  • Tests that don't exercise the change?
  • Secrets or overly broad permissions?

5. Verdict artifact

## Adversarial verdict
**Claims reviewed:** …
**Evidence:** files read · cmds + exit
**Confirmed:** …
**Refuted / gaps:** …
**Risk:** low|med|high
**Decision:** pass | fix-required | blocked

Done when

  • Independent reads + at least one automated check (or documented why impossible)
  • Explicit pass/fail with gaps list
  • No merge recommendation on testimony alone

Anti-patterns

¬Do
"LGTM" from PR textopen the files
same context that implementednew review role / session
ignore failing testsfail closed
duplicate full control-graph Outerlink CG; stay on evidence pack

Done_when cmds: git diff <base>...HEAD --stat · project verify cmds · verdict markdown present.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free